The strength of an open source project resides entirely in its developer community; a strong democratic culture of participation and hacking makes for a better piece of software. The key requirement is having developers who are not only willing to contribute, but also knowledgeable about the project’s internal structure and architecture. This session will introduce developers to the core internal architectural concepts of HBase, not just “what” it does from the outside, but “how” it works internally, and “why” it does things a certain way. We’ll walk through key sections of code and discuss key concepts like the MVCC implementation and memstore organization. The goal is to convert serious “HBase Users” into HBase Developer Users”, and give voice to some of the deep knowledge locked in the committers’ heads.

1. HBase Internals
Lars Hofhansl
Architect @ Salesforce.com
HBase Committer

2. HBase Internals
A.K.A. Get ready to:

3. Agenda
• Overview
• Scanning
• Atomicity/MVCC
• Updates
– Put/Delete
• Code!

4. A sparse, consistent,
distributed, multi-dimensional,
persistent, sorted map.

6. In the end it comes down to a
sorting problem.
How do you sort 1PB with 16GB
of memory?
756329184
756 329 184
567 239 148
123456789

7. Overview
• All changes collected/sorted in memory
• Periodically flushed to HDFS
– Writes to disk are batched
• HFiles periodically compacted into larger files
• Scanning/Compacting: Merge Sort

9. HDFS Directory Hierarchy
/hbase
/-ROOT-
/.META.
/.logs
/<table1>
/<region>
[/.recovered_logs]
/<column family>
/<HFile> Storage is per CF
/...
/...
/<column family>
/...
/<table2>
...

10. Scanning
• "Log Structured Mergetrees"
• Multiway mergesort
• Efficient scanning/seeking
/Region
|
|
+--/CF
|
|
+--/HFile

11. KeyValueHeap
• Maintains PriorityQueue of “lower” Scanners
• TopScanner in the Queue has the next KV

12. Scanning
• "Log Structured Mergetrees"
• Multiway mergesort
• Efficient scanning/seeking
/Region
|
|
+--/CF
|
|
+--/HFile

13. Updates
• All changes are written to the WAL first
• All changes are written to memory (the "MemStore")
• MemStores are flushed to disk (creating a new HFile)
• HFiles are periodically and asynchronously compacted into
fewer files.
• HFiles are immutable

14. ROW Atomicity
• Snapshot isolation and locking (per row)
o Row is locked while memory is changed
o Each row-update "creates" a new snapshot
o Each row-read sees exactly one snapshot
• Done with MultiVersionConcurrencyControl (MVCC)

15. Locking
• All KVs for a “row” are co-located in a region
• Locks are per row
• Stored in-memory at region server

16. MultiVersionConcurrencyControl
Wikipedia: "implement updates not by deleting an old piece of data
and overwriting it with a new one, but instead by marking the old
data as obsolete and adding the newer version"
Note that HBase calls it
MultiVersionConsistencyControl

17. MVCC writing
• Each write gets a monotonically increasing
"writenumber“
• Each KV written is tagged with this writenumber
(called memstoreTS in HBase)
• "Committing" a writenumber:
– wait for all prior writes to finish
– set the current readpoint to the writenumber

18. MVCC reading
• Each read reads as of a "readpoint“
– Filters KVs with a newer memstoreTS
• This is per regionserver (cheap in memory data
structures)
– if regionserver dies, current read is lost anyway
• but... "writenumbers" are persisted to disk
– for scanners that outlive a Memstore flush

20. MVCC
• Reader do not lock
• Transaction are committed strictly serially
• HBase has no client demarcated transactions
– A transaction does not outlive an RPC

21. Anatomy of an "update"
– Acquire a MVCC "writenumber"
– Lock the "row" (the row-key)
– Tag all KVs with the writenumber
– Write change to the WAL and sync to file system
– Apply update in memory ("Memstore")
– Unlock "row"
– Roll MVCC forward -> now change is visible
– If Memstore reaches configurable size, take a snapshot, flush it
to disk (asynchronously)

22. Anatomy of an "update"
Puts are optimized
– Acquire a MVCC "writenumber"
– Lock the "row" (the row-key)
– Tag all KVs with the writenumber
– Write change to the WAL and sync to file system
– Apply update in memory ("Memstore")
– Unlock "row"
o 5.5 sync WAL to HDFS without the row lock held
 If that fails undo changes in Memstore
 that works because changes are not visible, yet
– Roll MVCC forward -> now change is visible
– If Memstore reaches configurable size, take a snapshot, flush it
to disk (asynchronously)

23. Anatomy of an "update"
Puts are optimized, and batched
– Acquire a MVCC "writenumber"
– lock as many "rows" as possible
– Tag all KVs with the writenumber
– write all changes to the WAL and sync to file system
– apply all updates in memory ("Memstore")
– unlock all "rows"
o 5.5 sync WAL to HDFS without the row locks held
 If that fails undo changes in Memstore
 that works because changes are not visible, yet
– roll MVCC forward -> now changes are visible
– if Memstore reaches configurable size, take a snapshot, flush it
to disk (asynchronously)

24. Undo
• In-memory only
• Changes are not visible until MVCC is rolled
• Changes are tagged with the writenumber
• Undo removes KVs tagged with the writenumber

25. Deletes
• Nothing is deleted in place.
• a Delete sets a "tombstone marker" with a timestamp
• upon compaction deleted KeyValues are removed
• upon major compactions the tombstones are removed
• Three different tombstone marker scopes (all within a row)
o version - mark a specific version of a column as deleted
o column - mark all versions of a column as deleted
o column family - mark all versions of all columns of a column
family as deleted

26. Deletes, cont
• Delete markers always sort before KVs
• A scanner remembers markers it encounters
• Each KV is checked against remembered
markers
• Only one-pass is required for scanning
• Deletes are just KVs stored in HFiles

27. Let’s look at some code!

28. MVCC

29. MultiVersionConsistencyControl.java
public WriteEntry beginMemstoreInsert() {
synchronized (writeQueue) {
long nextWriteNumber = ++memstoreWrite;
WriteEntry e = new WriteEntry(nextWriteNumber);
writeQueue.add(e);
return e;
} Acquire a new Writenumber
}
public void completeMemstoreInsert(WriteEntry e) {
advanceMemstore(e);
waitForRead(e); Roll forward the readpoint
}
Wait for prior transactions to finish

30. MultiVersionConsistencyControl.java
boolean advanceMemstore(WriteEntry e) {
synchronized (writeQueue) {
e.markCompleted();
long nextReadValue = -1;
while (!writeQueue.isEmpty()) {
WriteEntry queueFirst = writeQueue.getFirst();
...
if (queueFirst.isCompleted()) {
nextReadValue = queueFirst.getWriteNumber();
writeQueue.removeFirst();
} else { Roll forward completed
break; transactions. Ordering is
} preserved.
}
...

31. MultiVersionConsistencyControl.java
...
if (nextReadValue > 0) {
synchronized (readWaiters) {
memstoreRead = nextReadValue;
readWaiters.notifyAll();
} Notify later transactions
}
if (memstoreRead >= e.getWriteNumber()) {
return true;
}
return false;
}
}

32. MultiVersionConsistencyControl.java
public void waitForRead(WriteEntry e) {
boolean interrupted = false;
synchronized (readWaiters) {
while (memstoreRead < e.getWriteNumber()) {
try {
readWaiters.wait(0);
} catch (InterruptedException ie) {
interrupted = true;
} Wait until write entry was
} applied.
}
if (interrupted) Thread.currentThread().interrupt();
}

33. Scanning

34. RegionScanner, creation
RegionScannerImpl(Scan scan,
List<KeyValueScanner> additionalScanners) {
...
IsolationLevel isolationLevel = scan.getIsolationLevel();
synchronized(scannerReadPoints) {
if (isolationLevel == IsolationLevel.READ_UNCOMMITTED) {
// This scan can read even uncommitted transactions
this.readPt = Long.MAX_VALUE;
MVCC.setThreadReadPoint(this.readPt);
} else {
this.readPt = MVCC.resetThreadReadPoint(mvcc);
}
scannerReadPoints.put(this, this.readPt);
} MVCC protocol
...

35. RegionScanner
Get all StoreScanners
...
for (Map.Entry<byte[], NavigableSet<byte[]>> entry :
scan.getFamilyMap().entrySet()) {
Store store = stores.get(entry.getKey());
StoreScanner scanner = store.getScanner(...);
scanners.add(scanner);
}
this.storeHeap = new KeyValueHeap(scanners, comparator);
}
Heap of StoreScanners

36. RegionScanner, cont.
public synchronized boolean next(
List<KeyValue> outResults, int limit) {
... MVCC
MVCC.setThreadReadPoint(this.readPt);
boolean returnResult = nextInternal(limit);
...
}
private boolean nextInternal(int limit) throws IOException {
if (isStopRow(currentRow)) {
...
return false;
} else { Get next KV from StoreScanners
byte [] nextRow;
do {
this.storeHeap.next(results, limit - results.size())
} while (Bytes.equals(currentRow, nextRow =peekRow()));
final boolean stopRow = isStopRow(nextRow);
}
...
}

37. StoreScanner
public synchronized boolean next(…) {
...
LOOP: while((kv = this.heap.peek()) != null) {
ScanQueryMatcher.MatchCode qcode = matcher.match(kv);
switch(qcode) {
case INCLUDE:
results.add(kv); Heap of Memstore/StoreFile scanners
this.heap.next();
... What to do with the KV:
continue;
case DONE_SCAN: ...
•Versions
case SEEK_NEXT_ROW: •TTL
reseek(matcher.getKeyForNextRow(kv));
break; •Deletes
case SEEK_NEXT_COL: ...
case SKIP:
this.heap.next();
break;
case SEEK_NEXT_USING_HINT:
KeyValue nextKV = matcher.getNextKeyHint(kv);
...
reseek(nextKV);
break;
...
}
}
}

38. Seek Hints
• Allow skipping many KVs without “touching”
• Seek-to(KV) instead of skip, skip, skip, …
• Used internally (for deletes, skipping older
versions, TTL)
• Used by filters

39. Memstore Scanner
public synchronized KeyValue next() {
if (theNext == null) {
return null;
}
KeyValueSkipListSet.iterator()
final KeyValue ret = theNext;
// Advance one of the iterators Snapshot during
if (theNext == kvsetNextRow) { flushes
kvsetNextRow = getNext(kvsetIt);
} else {
snapshotNextRow = getNext(snapshotIt);
}
// Calculate the next value
theNext = getLowest(kvsetNextRow, snapshotNextRow);
return ret;
}

40. Memstore Scanner
protected KeyValue getNext(Iterator<KeyValue> it) {
long readPoint = MVCC.getThreadReadPoint();
while (it.hasNext()) {
KeyValue v = it.next();
MVCC
if (v.getMemstoreTS() <= readPoint) {
return v;
}
}
return null;
}

41. Memstore Scanner
@Override
public synchronized boolean seek(KeyValue key) {
...
// kvset and snapshot will never be null.
// if tailSet can't find anything, SortedSet is empty (not
null).
kvTail = kvsetAtCreation.tailSet(key);
snapshotTail = snapshotAtCreation.tailSet(key);
return seekInSubLists(key);
}
For seek find the right tailSet
seekInSubLists() almost identical to next()

42. StoreFileScanner
private final HFileScanner hfs;
private KeyValue cur = null;
...
public KeyValue next() throws IOException {
KeyValue retKey = cur;
try {
// only seek if we aren't at the end.
// cur == null implies 'end'.
if (cur != null) {
hfs.next();
cur = hfs.getKeyValue();
skipKVsNewerThanReadpoint();
}
} catch(IOException e) {
throw new IOException("Could not iterate " + this, e);
}
return retKey;
}

43. HFileScanner/Reader
@Override
public boolean next() throws IOException {
...
blockBuffer.position(...);
...
if (blockBuffer.remaining() <= 0) {
long lastDataBlockOffset = Still on current block?
reader.getTrailer().getLastDataBlockOffset();
// read the next block
HFileBlock nextBlock = readNextDataBlock();
if (nextBlock == null) {
return false;
}
updateCurrBlock(nextBlock); If not, read the next block
return true;
}
// We are still in the same block.
readKeyValueLen();
return true; Mark the next KV in the buffer
}

44. Puts

45. Batch Put
• private long doMiniBatchPut(BatchOperationInProgress<…> batchOp){
WALEdit walEdit = new WALEdit();
...
MultiVersionConsistencyControl.WriteEntry w = null;
...
try {
// STEP 1. Try to acquire as many locks as we can
// STEP 2. Update any LATEST_TIMESTAMP timestamps
Begin transaction
// Acquire the latest mvcc number
w = mvcc.beginMemstoreInsert();
// STEP 3. Write back to memstore
for (int i = firstIndex; i < lastIndexExclusive; i++) {
addedSize += applyFamilyMapToMemstore(familyMaps[i], w);
}
// STEP 4. Build WAL edit
for (int i = firstIndex; i < lastIndexExclusive; i++) {
addFamilyMapToWALEdit(familyMaps[i], walEdit);
}

46. Batch Put, cont.
// STEP 5. Append the edit to WAL. Do not sync wal.
txid = this.log.appendNoSync(regionInfo, …, walEdit, …);
// STEP 6. Release row locks, etc.
this.updatesLock.readLock().unlock(); Write WAL record
for (Integer toRelease : acquiredLocks) { but don’t sync!
releaseRowLock(toRelease);
} Guard against concurrent flushes
// STEP 7. Sync wal.
this.log.sync(txid);
walSyncSuccessful = true; Sync after locks are released
// STEP 8. Advance mvcc.
// This will make this put visible to scanners and getters.
if (w != null) {
mvcc.completeMemstoreInsert(w);
w = null;
}
Commit
...
return addedSize;
}

47. Batch Put, something went wrong
} finally {
if (!walSyncSuccessful) {
rollbackMemstore(batchOp, familyMaps,
firstIndex, lastIndexExclusive);
}
if (w != null) mvcc.completeMemstoreInsert(w);
if (locked) {
this.updatesLock.readLock().unlock();Always complete
} the transaction!
for (Integer toRelease : acquiredLocks) {
releaseRowLock(toRelease);
}
...
}

48. Memstore changes
private long applyFamilyMapToMemstore(
Map<byte[], List<KeyValue>> familyMap,
WriteEntry localizedWriteEntry) {
long size = 0;
boolean freemvcc = false; Can pass a write entry
that spans mutliple calls
try {
if (localizedWriteEntry == null) {
localizedWriteEntry = mvcc.beginMemstoreInsert();
freemvcc = true;
}
for (Map.Entry<…> e : familyMap.entrySet()) { This begins the
byte[] family = e.getKey(); transaction
List<KeyValue> edits = e.getValue(); (MVCC)
...

49. Memstore changes
Tag KV with write number (MVCC)
...
Store store = getStore(family);
for (KeyValue kv: edits) {
kv.setMemstoreTS(localizedWriteEntry.getWriteNumber());
size += store.add(kv);
}
}
} finally {
if (freemvcc) {
mvcc.completeMemstoreInsert(localizedWriteEntry);
}
}
return size;
} This makes the changes visible

50. Deletes

51. ScanDeleteTracker (checking markers)
public void add(buffer, qualifierOffset, qualifierLength, ts, type) {
if (!hasFamilyStamp || ts > familyStamp) {
if (type == KeyValue.Type.DeleteFamily.getCode()) {
hasFamilyStamp = true;
familyStamp = ts; Only remember TS for family deletes
return;
}
if (deleteBuffer != null && type < deleteType) {
// same column, so ignore less specific delete
if (Bytes.equals(deleteBuffer, deleteOffset, deleteLength,
buffer, qualifierOffset, qualifierLength)){
return;
} A version delete
} marker can be
// new column, or more general delete type
deleteBuffer = buffer; ignored if there is a
deleteOffset = qualifierOffset; column marker
deleteLength = qualifierLength;
deleteType = type; already.
deleteTimestamp = ts;
}
Remember the KV
}

52. ScanDeleteTracker (checking markers)
public DeleteResult isDeleted(buffer, qualifierOffset,qualifierLength,
timestamp) {
if (hasFamilyStamp && timestamp <= familyStamp) {
return DeleteResult.FAMILY_DELETED; Family marker case
}
if (deleteBuffer != null) {
int ret = Bytes.compareTo(deleteBuffer, deleteOffset, deleteLength,
buffer, qualifierOffset, qualifierLength);
if (ret == 0) {
if (deleteType == KeyValue.Type.DeleteColumn.getCode()) {
return DeleteResult.COLUMN_DELETED;
}
Column marker case
// If the timestamp is the same, keep this one
if (timestamp == deleteTimestamp) {
return DeleteResult.VERSION_DELETED;
} Version marker case
// different timestamp, let's clear the buffer.
deleteBuffer = null;
} else if(ret < 0){ HFiles scanned newest
// Next column case. TS first
deleteBuffer = null;
} else {
throw new IllegalStateException(...);
}
}
return DeleteResult.NOT_DELETED;
}

53. Questions?
Comments?
More details on http://hadoop-hbase.blogspot.com